PREDICTIVE CUSTOMER ANALYTICS

Peter Fader
Frances and Pei-Yuan Chia Professor of Marketing
Co-Director, Wharton Customer Analytics Initiative
The Wharton School of the University of Pennsylvania
http://wcai.wharton.upenn.edu
Twitter: @faderp

WHARTON ONLINE

SETTING THE STAGE

Organiza(on:

A public radio sta/on supported primarily by contribu/ons

from its listeners

Challenge:

Looking at listeners histories of whether or not they gave

each year, what can we predict about their future giving
pa=erns?

Focal donors:

Reference:

- Ini/al focus on 1995 cohort, ignoring dona/on amount

- 11,104 rst-/me supporters who made a total of 24,615
repeat dona/ons over the next 6 years
Fader, Peter S., Bruce G.S. Hardie, and Jen Shang (2010),
Customer-Base Analysis in a Discrete-Time Noncontractual
SeYng, Marke&ng Science, 29 (6), 1086-1108.

WHARTON ONLINE

Predic/ve Analy/cs

HOW MUCH WILL DONORS GIVE IN THE FUTURE?!

HOW DOES IT DEPEND ON THEIR PAST PATTERNS?!

ID

1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

100001

100002

100003

100004

100005

100006

100007

100008

100009

100010

111102

111103

111104

...

WHARTON ONLINE

Predic/ve Analy/cs

L E T S F I R S T L O O K AT B O B : !
W H AT C A N W E P R E D I C T A B O U T H I S G I V I N G I N 2 0 0 2 - 0 6 ? !

ID

1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

100001

100002

100003

100004

100005

100006

100007

BOB

100009

100010

...

WHARTON ONLINE

Predic/ve Analy/cs

W H AT C A N W E T E L L A B O U T S A R A H ? !

ID

1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

SARAH

100002

100003

100004

100005

100006

100007

BOB

111102

111103

111104

...

WHARTON ONLINE

Predic/ve Analy/cs

HOW DO MARY AND SHARMILA COMPA RE?

ID

1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

SARAH

100002

100003

MARY

100005

100006

100007

BOB

SHARMILA 1

100010

...

WHARTON ONLINE

Predic/ve Analy/cs

W H AT D O N AT I O N B E H AV I O R C H A R A C T E R I S T I C S !
D O W E N E E D T O TA K E I N T O A C C O U N T ?
G I V I N G B E H AV I O R S

D O N O R T Y P E S

How recently did the

donor give? When was
the last /me the donor
gave?

R E C E N C Y:

F R E Q U E N C Y: How many /mes did the

donor give in the past 6
years?

ALIVE:

Clearly an ac/ve donor:

giving frequently and
gave recently

D O R M A N T:

Has not given recently,

but is likely to give
again with the right
development prompts

LAPSED:

Has not given recently

and is not likely to give
again

Most meaningful dona/on pa=erns can be

described by these two metrics alone.

ID
1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

MARY

SHARMILA 1

WHARTON ONLINE

Predic/ve Analy/cs

MORE ABOUT RECENCY AND FREQUEN CY

Y Y N N N N
Y Y N N Y Y

What does it mean when theres one or more no dona(on at the end of a
sequence?
a) The donor lapsed (i.e., le` the donor pool)
b) The donor is dormant (i.e., decided not to give that year, didnt think of
giving, etc.)
c) We dont know, but can build a model to come up with a best guess

Answer:
c) We never know for sure whether the donor is lapsed or not; based
on recency and frequency of his dona/on, we can make an educated guess
about the probability of lapsing, so we can decide where to devote resources

Based on our best guesses about the probability of death and propensity to
donate, we can calculate expected frequency of future dona(ons for each
donor
WHARTON ONLINE

Predic/ve Analy/cs

W H AT A B O U T M A RY V E R S U S C H R I S ? !

ID

1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

SARAH

100002

100003

MARY

100005

100006

100007

BOB

111102

CHRIS

111104

...

WHARTON ONLINE

Predic/ve Analy/cs

You can try this at home

WHARTON ONLINE

BUY TILL YOU DIE MODEL

We employ a Buy Till You Die model to predict future dona/on behaviors

The model only uses three inputs:
1. Recency (R)
2. Frequency (F)
3. Number of people for each combina/on of R/F
This requires a small amount of data and provides an easier structure to work
with (i.e., data are aggregated from individual-level to R/F groups)

By assuming certain probability distribu(ons for donors propensi(es, we can
construct a robust model that is easy to implement on Excel

This BTYD modeling approach has a long track record of success in a variety of
dierent domains

WHARTON ONLINE

Predic/ve Analy/cs

EXCEL IMPLEMENTATION

Detailed step-by-step instruc/ons available at h=p://brucehardie.com/notes/010/

WHARTON ONLINE

Predic/ve Analy/cs

E X P E C T E D # O F D O N AT I O N S I N 2 0 0 2 - 2 0 0 6 !
AS A FUNCTION OF RECENCY AND FREQUENCY

Name

1995

1996

1997

1998

1999

2000

2001

BOB

SARAH

MARY

SHARMILA

CHRIS

# Rpt Trans
(1996-2001)

1995

0.07

1
2
3
4
5
6
WHARTON ONLINE

Year of Last Transac(on

1996

1997

1998

1999

2000

2001

0.09

0.31

0.59

0.84

1.02

1.15

0.12

0.54

1.06

1.44

1.67

0.22

1.03

1.80

2.19

0.58

2.03

2.71

1.81

3.23
3.75

Predic/ve Analy/cs

ANALYSIS
Bob (R:6, F:6) is expected to donate 3.75 /mes out of 5 opportuni/es between 2002 and
2006, surprisingly low given his 100% dona/on rate
Mary and Chris have the same RF (6,4), so their expected number of dona/ons going
forward is the same
Even though Mary and Chris have lower F than Sharmila, their higher R suggests that
they are Alive, thus they are 50% more valuable than Sharmila
Sharmila (5,5), despite high dona/on rate, has likely lapsed
Sarah, with very low R and F, is lapsed and/or a very light donor (hard to tell)









WHARTON ONLINE

# Rpt Trans
(1996-2001)
0
1
2
3
4
5
6

Year of Last Transac(on

1995

1996

1997

1998

1999

2000

2001

0.09

0.31

0.59

0.84

1.02

1.15

0.12

0.54

1.06

1.44

1.67

0.22

1.03

1.80

2.19

0.58

2.03

2.71

1.81

3.23

0.07

3.75

Predic/ve Analy/cs

HOW DO WE KNOW THE MODEL WORK S?

The model does a solid job of making predic(ons about future dona/on
behaviors, as condi(onal expecta(ons correspond very well with actual holdout
period data (2002 2006)

WHARTON ONLINE

Predic/ve Analy/cs

HOW DO WE KNOW THE MODEL WORK S?

Overall, the model is excep/onal good at bng the historical data as well as
the holdout period data (i.e., model predic/ons vs. actual number of dona/ons
between 2002 and 2006)

WHARTON ONLINE

Predic/ve Analy/cs

A S E C O N D I L L U S T R AT I O N


Using a larger dataset from a dierent non-prot rm, we create a heat map that
shows which combina/ons of RF will likely yield the most valuable donors

WHARTON ONLINE

1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
x / tx
0
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19
0
0.01
1
0.00 0.01 0.01 0.02 0.03 0.04 0.06 0.07 0.09 0.11 0.13 0.15 0.17 0.19 0.22 0.24 0.26 0.29 0.31
2
0.00 0.00 0.01 0.02 0.03 0.05 0.07 0.09 0.13 0.16 0.20 0.25 0.29 0.34 0.39 0.43 0.47 0.51
3
0.00 0.00 0.01 0.01 0.03 0.05 0.07 0.11 0.16 0.22 0.29 0.37 0.44 0.52 0.59 0.66 0.71
4
0.00 0.00 0.00 0.01 0.02 0.05 0.08 0.13 0.21 0.30 0.41 0.52 0.64 0.74 0.84 0.91
5
0.00 0.00 0.00 0.01 0.02 0.05 0.09 0.17 0.28 0.41 0.57 0.74 0.89 1.01 1.12
6
0.00 0.00 0.00 0.01 0.02 0.06 0.12 0.23 0.39 0.59 0.81 1.02 1.19 1.32
7
0.00 0.00 0.00 0.01 0.03 0.08 0.17 0.33 0.58 0.86 1.14 1.36 1.52
8
0.00 0.00 0.00 0.01 0.04 0.11 0.26 0.52 0.88 1.23 1.52 1.72
9
0.00 0.00 0.00 0.02 0.06 0.19 0.44 0.85 1.31 1.68 1.92
10
0.00 0.00 0.01 0.03 0.12 0.34 0.79 1.36 1.83 2.12
11
0.00 0.00 0.01 0.06 0.24 0.68 1.37 1.97 2.32
12
0.00 0.00 0.03 0.15 0.54 1.33 2.10 2.53
13
0.00 0.01 0.07 0.39 1.23 2.20 2.73
14
0.00 0.03 0.23 1.06 2.28 2.93
15
0.01 0.11 0.81 2.30 3.13
16
0.03 0.51 2.25 3.33
17
0.20 2.04 3.53
18
1.48 3.73
19
3.94

Predic/ve Analy/cs

PROBABILITY OF BEING ALIVE

x/tx

1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009
0

0.06

0.01 0.03 0.05 0.07 0.10 0.13 0.17 0.22 0.27 0.32 0.38 0.45 0.51 0.58 0.65 0.73 0.80 0.87 0.93

0.00 0.01 0.02 0.03 0.06 0.09 0.12 0.17 0.23 0.30 0.37 0.45 0.54 0.62 0.70 0.79 0.86 0.93

0.00 0.00 0.01 0.02 0.03 0.06 0.10 0.15 0.21 0.29 0.38 0.48 0.58 0.68 0.77 0.86 0.93

0.00 0.00 0.00 0.01 0.02 0.05 0.08 0.14 0.21 0.31 0.42 0.53 0.65 0.76 0.85 0.93

0.00 0.00 0.00 0.01 0.02 0.04 0.08 0.14 0.23 0.35 0.48 0.62 0.74 0.85 0.93

0.00 0.00 0.00 0.01 0.02 0.04 0.09 0.16 0.28 0.42 0.58 0.72 0.84 0.93

0.00 0.00 0.00 0.01 0.02 0.05 0.10 0.21 0.35 0.53 0.70 0.84 0.93

0.00 0.00 0.00 0.01 0.02 0.06 0.14 0.28 0.48 0.67 0.83 0.93

0.00 0.00 0.00 0.01 0.03 0.09 0.22 0.41 0.64 0.82 0.93

10

0.00 0.00 0.00 0.01 0.05 0.15 0.35 0.60 0.81 0.93

11

0.00 0.00 0.00 0.03 0.10 0.27 0.55 0.79 0.93

12

0.00 0.00 0.01 0.05 0.20 0.49 0.78 0.93

13

0.00 0.00 0.03 0.13 0.42 0.75 0.93

14

0.00 0.01 0.07 0.34 0.73 0.93

15

0.00 0.03 0.24 0.69 0.93

16

0.01 0.14 0.63 0.93

17

0.05 0.54 0.93

18

0.37 0.93

19

0.93

WHARTON ONLINE

Predic/ve Analy/cs

SUMMARY: HOW IS THIS METHOD DIFFERENT?

There are many models that predict future dona/on behaviors; we believe our
method is dierent / superior because:

1. The model requires a very small amount of data (Recency and Frequency),
compared to other models that require a large dataset (typically detailed
individual-level characteris/cs, e.g., demographics)
2. The model has demonstrated robust out-of-sample valida(on
3. The model can be generalized to other types of behaviors; it is not
excessively customized to the dona/on domain
4. The model can easily be implemented on Excel; it does not require any
proprietary or specialized so`ware


WHARTON ONLINE

Predic/ve Analy/cs

WA N T M O R E ?
For large-scale databases, use our open-source BTYD R Library:
h=p://cran.r-project.org/web/packages/BTYD/BTYD.pdf


While this model oers accurate predic/ons and useful insights about how to
understand dona/on propensi/es, it stops short of oering any specic advice
about which donors to target and when / how to do so

Building on this model, Schweidel and Knox explore the impact of fundraising
eorts on dona/on ac/vity in: Incorpora/ng Direct Marke/ng Ac/vity into
Latent A=ri/on Models (h=p://ssrn.com/abstract=1670060)





WHARTON ONLINE

Predic/ve Analy/cs

DISCUSSION

Professor Peter Fader
www.petefader.com
Twi=er: @faderp

Wharton Customer Analy/cs Ini/a/ve
www.wharton.upenn.edu/wcai/

WHARTON ONLINE

Predic/ve Analy/cs